[FEA]: Expose PDL option on the tuning side
- Dominant language
- C++
- Stars
- 2.5k
- Forks
- 486
- Avg merge
- 2d 6h
- Merged PRs (30d)
- 295
Description
In https://github.com/NVIDIA/cccl/pull/9247 we introduced PDL to `DeviceRadixSort`. We observed that for some workloads PDL was benefiting performance but for others it was regressing and thus we chose to enable it conditionally. This condition is internal. Given that there could be other cases were PDL will either benefit or even by accident regress the performance for the users, users should be able to toggle PDL on and off. In other words we should lift the internal condition for PDL on the API side uniformly for all algorithms that use PDL.
- [ ] Figure a design and expose PDL on the tuning side for `DeviceRadixSort`
- [ ] See if other algorithms are impacted and open corresponding sub-issues
ping @bernhardmgruber @gevtushenko for viz once back
Contributor guide
Assessment
This issue has not been assessed yet.